論文ID: 25-00040
Adaptive cruise control (ACC) is one of the critical elements of vehicle performance in the market. To ensure the quality of ACC performance, comprehensive evaluations that control both complex test scenarios that reproduce market driving conditions and vehicle behavior is required. However, it is difficult to evaluate all combinations of test scenarios using real test vehicles within limited development resources. Furthermore, it is necessary to determine complex Electrical Control Unit (ECU) parameters while considering multiple performance trade-offs. This paper proposes a new automatic screening and exploration system for ACC, incorporating Bayesian active learning (BAL). The proposed system automatically explores the worst conditions of ACC and the design space of ECU parameters for the improvement of vehicle ACC performance. This system consists of two automated elements: an automatic evaluation system and an automatic exploration system. In the automatic evaluation system, the behavior of ACC is automatically evaluated in real-time simulation using Real Car Simulation Bench (RC-S). Additionally, ACC sensor simulation is used to simulate various driving scenarios that may occur in the market. In the automatic exploration system, the worst condition screening evaluation of ACC performance and the exploration of the feasible region of design space for ECU parameters using BAL are conducted. As a result, it becomes possible to make the evaluation process more efficient through closed-loop evaluation, thereby improving ACC performance. In BAL, a Gaussian process model of ACC performance evaluated by RC-S is trained. Based on the posterior distribution of the trained Gaussian process model, the acquisition function is evaluated and maximized to generate new sampling points. In this study, an example of data comparison between RC-S and a real vehicle driving on a test course is demonstrated to show the effectiveness of the proposed system.
